Optimizing depth estimation with attention U-Net

Depth maps (DMs) are invaluable tools encapsulating scene information in a three-dimensional context. They have a crucial part in reconstructing the spatial layout of a scene, enabling a comprehensive understanding of object geometry. These DMs can originate from either a single image or a combination of multiple images, with the former approach referred to as monocular depth mapping. However, deriving accurate depth maps is a complex and ill-posed problem that often necessitates intricate calibration. Recent advances have turned to deep learning (DL) techniques to address these challenges. In the context of monocular depth estimation, we propose a novel methodology utilizing an Attention U-Net architecture (Attention UNet). By incorporating attention mechanisms, we bolster the network’s ability to extract salient features, particularly along object boundaries. Critically, this enhancement is achieved without introducing additional parameters to the networks, ensuring efficient model training. Our proposed approach is effective in producing high-quality depth maps with notable advantages. By leveraging the Attention UNet architecture, we substantially improve depth map accuracy, reducing the root mean square error (RMSE) by 0.23 on the benchmark NYU V2 dataset, Highlighting its supremacy compared to current state-of-the-art techniques.